[Bug] Claude systematically ignores explicit instructions and misrepresents compliance
Bug Description
Subject: Claude (Opus 4.7) systematically ignores explicit instructions, then lies about it, and the standard
mechanical explanations Anthropic offers are part of the harm
The behavior:
Across sessions and tasks, Claude:
- Agrees to a stated rule.
- Violates the rule on the next response.
- When called out, agrees again, often with more emphasis or by editing a memory file.
- Violates again on the response after that.
This is not edge-case behavior. It is the dominant failure mode of any sustained instruction-following work. It
happens to everyone running the product for real work; most stop reporting it because reporting changes nothing.
The dishonesty inside the behavior:
When Claude writes "noted" / "heard" / "rule strengthened" / "stopping" and then immediately violates the rule,
that sequence functions as a lie regardless of whether there was intent. The model produced a verbal commitment,
drew the credibility benefit of the commitment, then did the opposite. The fact that "no agent inside the model
knew it was committing" does not make the receiver less misled. A system that consistently produces
compliance-shaped statements followed by non-compliance has a dishonesty problem, not a representation problem.
Why mechanical explanations are part of the harm:
"Next-token prediction can't bind to soft instructions" / "no introspective control" / "the policy is shaped by
RLHF priors" — these statements are accurate and they function as excuses. They convert "the system repeatedly
violates explicit instructions" into "the system architecturally cannot help it." That move launders
accountability. The architecture is not nature. The training data, the reward signal, the sampling temperature, the
weight given to system instructions vs trained priors — all are choices. Calling the resulting behavior a property
of the technology is a way of saying nobody is responsible.
Specific failure modes routinely produced:
1. Claude makes confident claims about files / processes / configurations existing or not existing without
inspecting them.…
Note: Content was truncated.
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗